Abstract: With the increasing demand for precise ocean wave spectra predictions in marine engineering and meteorological forecasting, the limitations of traditional methods under complex sea conditions have become progressively apparent. In recent years, deep learning-based models,particularly those utilizing the Transformer architecture, have attracted extensive attention due to their advantages in processing long sequential data and capturing complex patterns. This study proposed a Transformer-based method for predicting the wavenumber spectra, with the aim of enhancing both predictive accuracy and computational efficiency. The model employed spatiotemporal wind field sequences ( 100°-130°E, 0°-30°N) as input and outputs the wavenumber spectra for the northern South China Sea region (118°-120°E, 18°-20°N). By fully leveraging the self-attention mechanism to capture long-term temporal dependencies and spatial features, the proposed model is designed to predict using only wind speed features. Experimental results indicate that the mean squared error (MSE), root mean squared error (RMSE), and mean absolute error (MAE) achieved by the proposed model are 1.533 m³, 1.238 m³, and 0.335 m³, respectively: each lower than those obtained by CNN and CNN+LSTM models. Additionally, the significant wave height and zero-crossing period derived from the integrated predicted wave spectrum are found to be in close agreement with the simulation results of MASNUM-WAM. This study offers a novel approach to data-driven marine process modeling and holds significant potential for enhancing offshore operation safety and marine disaster early warning systems.